Methodology

Every price here comes from a dealer listing or a published auction result. This page sets out how those become the curves and indices, and where they fall short.

The data

We track 6,336 specimens across 18 fossil types, carrying 19,450 observed prices from 1997-10-04 to 2026-08-16. 19,441 are retail (dealer listings and marked-sold items), 9 are auction results.

Dealer sites are scraped for current listings. History predating our own tracking comes from their archived pages on the Internet Archive. Scraping respects robots.txt, with two to three seconds between requests. Auction results are entered by hand from published sale records, since the major houses can't be scraped.

Asking prices, mostly

Most of what we track is an asking price. A listing records what a dealer wanted for a piece, which sits above what it eventually brought. Only auction results and dealer-marked sold items are completed transactions. Read the curves as the shape of the retail market; a quick private sale usually lands under the low end of the range.

The valuation model

For each type we fit price against size by weighted least squares in log space:

ln(price) = b₀ + b₁·ln(size) + b₂·ln(size)² + Σ condition·δ + Σ element·δ + Σ variety·δ + Σ form·δ

The squared term lets the size exponent grow with size, which a single power law can't do. For megalodon teeth the local exponent runs from about 3.0 at 3.76 inches to 4.8 at 5.74 inches. Past the vertex of the fitted parabola the curve would bend back down, so size is clamped to the monotonic side and the curve stays non-decreasing.

Observations carry a two-year half-life, so the curve follows the current market rather than averaging every year we've tracked. The band around each estimate is the weighted RMS of the log residuals: the spread of actual specimens around the curve rather than the standard error of the fit, so it's wide, covering roughly the middle two-thirds of the market at that size.

A type needs 15 sized observations before we fit it. Below that, no curve is published.

What the model controls for

Beyond size, five covariates, each estimated from the data:

• Condition — partial, repaired, restored, composite, parsed from the listing title. A flag enters the regression once five examples exist in the bucket.
• Element — jaw, vertebra, limb, skull, rib, osteoderm, frill, horn, toe.
• Variety — ammolite and iridescent nacre against plain ammonite; genus for trilobites.
• Form — plates, clusters and pairs against single specimens.
• Size — parsed from titles, normalised to inches.

Without them, a bucket's average moves with its composition and reads as a price change.

The price index

The Evolution tab measures change rather than level. Hedonic regression with time-period dummies:

ln(price) = b·ln(size) + Σ condition·δ + Σ element·δ + Σ variety·δ + Σ form·δ + Σ period·γ_t

The period coefficients γ_t give the price level holding size and quality fixed, indexed to 100 at each series' first period. The same coefficients produce the standard-specimen view: the fitted price of a median-sized clean example, quarter by quarter. Dollars-per-inch, which this replaced, tracked size mix as much as price.

Most series are quarterly. Thin but genuinely traded ones fall back to annual buckets, marked in the legend.

What gets excluded

Jewellery, replicas, display stands, cases, books and gift items are filtered out by title. Items without a parseable size stay in the history but sit out the size models. Titles that contradict their category get reclassified: mosasaur jaw sections had been sitting in the tooth curve, dragging it upward.

“On the market now” means the specimen appeared in its own source's most recent scrape. Anything older is history, which keeps items sold between weekly scrapes from showing as available.

Where this is weak

• Pre-2023 history reads high. The archived pages behind it mostly captured the first page of each category, sorted by price. The model can't see the sampling that caused it.
• “(various)” buckets stay noisy. Mixed-species groups keep within-bucket variation the covariates can't reach, so comparisons default to single-species series.
• Auction figures include buyer's premium, some converted at approximate rates. Headline lots represent themselves only.
• Locality and provenance aren't modelled yet, so formation premiums land in the residual.
• Thin series are indicative. Twenty observations give a sketch.

How current this is

Sources are re-scraped weekly, followed by a sweep that flags implausible sizes, outliers against the fitted curve, and accessories that slipped the filter. Latest observation: 2026-08-16.

Both models are least squares, solved directly. Every page links the listings behind its numbers.